Thanks for this! Have you considered hosting your code on google code or sourceforge? Seems to me it would be easier for both you and everyoe else that way...

Yes I will do all of the above, once I am "done". I don't want to publish unfinished code. I put it up here so I can get feedback on what I need to do to make it perfect. So when we have it bronzed there are no defects.

Before I put it out for wide distribution I would like to let it settle for a couple of weeks and give people a chance to play with it. And also, independent confirmation of the filter's output (by measuring the magnitude and phase response).

I'm working on SSE3 optimized version of CascadeFilter, its looking promising.

No, sorry to say. A filter should produce bounded output for any bounded input.

What if I'm using a Biquad that has a bit of resonance from the Q? If you filter the same block of noise over and over again it will get high values in the resonating part of the frequency spectrum, I think...

When I lowered the Q the problem went away

File is updated:

- Added optimized template specialization for CascadeFilter::ProcessI()
in SSE3 intrinsics (MSVC) when channels=2 and stages=1 (all Biquads)

Sorry for the compiler specific stuff. If someone wants to help me add the necessary stuff to make it work with compilers other than MSVC, I would gladly add it in.

For the case of Biquads, and any Butterworth or Chebyshev filters with poles=2 channels=2, the SSE3_OPTIMIZED version is 38% faster.

If the Q value was the cause and you lowered the Q, great. The fact remains that a useful filter gives bounded output no matter what bounded input it receives.

I agree but now that I think about it, the filter's input was unbounded. Because I was passing the same buffer back in over and over - each successive pass kept raising the value until it overflowed.

Wow, this code looks cool.
But since I don't understand those depths of dsp programming, I'm sorry I can't say anything about the quality of the code itself, but how it is documented and how it is written (I mean it's style) is great.
One thing I noticed is the use of the postfix decrement and increment operators. Whenever possible, you should use the prefix versions of them for better efficiency. Most modern compilers will optimize the postfix versions of them under specific circumstances, but to be sure you should go for the prefix versions yourself. Especially the for loops seem to be predestinated for this little change, some other code must not be changed for sure, since it relies on the not yet incremented/decremented value of the variables.
If you do not already know, see e.g. here
http://www.thunderguy.com/semicolon/2002/08/13/prefer-prefix -operators-over-postfix/
for a more detailed explanation.

Keep up the good work,
Patrik

Do you mean that you were feeding the output data back into the input? That can cause a perfectly good filter to overflow. On the other hand, if you keep your input and output buffers separate, never changing your input buffer data, feeding it to the filter repeatedly, and the filter overflows, you have a problem.

Yes that is what I was doing. I wrote a loop to time the performance and it filtered the same buffer repeatedly. Eventually as you said, it overflowed.

I'm not sure I'm following the conversation here, because it morphed from denormals to overflow... unless I'm missing something in which case just ignore me. But one thing I learned recently is that filters often use internal saturation to prevent 'overflow' and keep things stable under all conditions. The result in your case (the feedback loop) will be a ringing at the center frequency.

Updated the file:

- Added SSE3 optimization for a general number of stages in
ProcessI() when channels==2

The general case of N stage 2 channel filtering in ProcessI() is now 65% faster with SSE3 instructions.

Updated the file. Provided lots of examples (see the function at the bottom) on how to use the classes for template-impaired folks (which I was not too long ago).

I'll provide feedback as soon as I finish my oscillators, thanks for the stuff.

SourceForge page:

http://sourceforge.net/projects/dspfilterscpp/